BUG: expanding/rolling_quantile() inconsistent with Series.quantile()

[seth-p]

It appears that Series.quantile() interpolates between the two bracketing values, whereas expanding/rolling_quantile() don't.

In [496]: x = Series([0,1])

In [497]: x.quantile(0.3)
Out[497]: 0.29999999999999999

In [498]: expanding_quantile(x, 0.3)
Out[498]:
0    0
1    0
dtype: float64

In [499]: rolling_quantile(x, 2, 0.3, min_periods=1)
Out[499]:
0    0
1    0
dtype: float64

[seth-p]

My 2c: there really shouldn't be two different implementations of a quantiles functions.

[TomAugspurger]

Looks like either way is mathematically valid, but I agree we should be consistent.
It's probably best to add a keyword to one of the implementations.

I'm looking at the numpy percentile function (which we use in DataFrame.quantile), and it mentions

The values and distances of the two
nearest neighbors as well as the interpolation parameter will
determine the percentile if the normalized ranking does not match q
exactly.

But it looks like this is parameter was only added in numpy 1.9 (which I think is the dev version).

numpy 1.8 docs: http://docs.scipy.org/doc/numpy-1.8.1/reference/generated/numpy.percentile.html#numpy.percentile
numpy 1.9 docs: http://docs.scipy.org/doc/numpy-dev/reference/generated/numpy.percentile.html

Anyway right now you can get the same behavior with df.interpolate(method='nearest') as long as you have scipy installed.

[TomAugspurger]

I share your desire that rolling/expanding/regular should share an implementation, but NDFrame.interpolate allows for using a bunch of scipy methods. I'm not sure if we could do that in the Cython code.

[jreback]

dupe of #9413